Official announcement of Vstorm and Pydantic partnership

Vstorm's partnership with Pydantic began with a decision to build on an AI framework before most teams had heard of it, and to contribute fixes and tooling back to the codebase when production gaps appeared. This article tells that story: from early adoption in beta, through 30+ client deployments, to open-source contributions recognised by the Pydantic team, to the formalised partnership announced today in this article, including Vstorm's role as implementation partner and ambassador.

Some partnerships are announced before the work begins, but this is quite the opposite. By the time we at Vstorm and our partners at Pydantic formalised our relationship, we had already spent more than a year building on the framework in production, contributing fixes to the core codebase, and shipping five open-source tools to address gaps that client deployments had exposed. The announcement is the end of one story and the start of another.

The beginning: building on the framework before it had a version number #

We started using Pydantic AI during its beta period, before v1.0 existed and before most engineering teams had considered it as a production option. That decision was not speculative. It came from familiarity with the underlying Pydantic validation library, which is embedded in the OpenAI SDK, the Anthropic SDK, LangChain, LlamaIndex, and AutoGPT, among others. By early 2026, the library had reached 550 million monthly downloads and 10 billion total downloads, with active adoption at all FAANG companies and 20 of the 25 largest companies on NASDAQ. (Source: pydantic.dev)

When the Pydantic team launched an agent framework built on the same foundation, we read the design choices and recognised the same discipline that made the validation library trustworthy: full type safety, model-agnostic architecture, and first-class OpenTelemetry instrumentation that feeds straight into Pydantic Logfire, the team's AI observability platform. These were the properties we needed for production-grade agentic AI work. We started building on it.

The journey: taking it into real client deployments #

Through 2024 and into 2025, while Pydantic AI was starting to expand to the wider market, we were already running it inside client systems.

The Hybrid Agent-Graph Architecture case study documents one of those early engagements: the progression from a single-agent system to a hybrid architecture using Pydantic AI production agents with Text-to-SQL capabilities. The case study is public and specific. It shows the production requirements that exposed the limits of simpler architectures; state management, multi-step reasoning, structured database queries; and how the right design addresses each constraint.

For Mixam, a global self-publishing platform, we built an agent to guide customers through complex printing specifications in real time: layered domain knowledge, multi-step product logic, and a live customer interaction surface. The system reduced the support burden on the Mixam operations team while improving the ordering experience for users who needed detailed guidance.

For the ARIJ Network, connecting investigative journalists across 22 Arabic-speaking countries, we built a bilingual RAG-based chatbot that moved their training delivery from handling 1% of enquiries manually to providing reliable, fact-checked knowledge at scale. That system also runs on Pydantic AI.

By the time Pydantic AI reached v1.0 in September 2025, we had already shipped it across multiple production environments.

The battles: contributing back when production exposed gaps #

Building on a beta framework in production means encountering problems the framework had not yet solved. Rather than working around those gaps with client-specific patches, we addressed them at the source.

Vstorm is listed among the core contributors to Pydantic AI. The full record of our pull requests to the main repository is public: github.com/pydantic/pydantic-ai/pulls. These contributions reflect specific problems we encountered in client work and solved at the framework level, so that other teams building in the same space would not have to solve them again.

Alongside those contributions to the core codebase, we built and open-sourced five tools that address the recurring engineering problems that appear when moving from prototype to deployed system:

• pydantic-deep — extends Pydantic AI with planning, sandboxed code execution, subagent delegation, skills, checkpoints, and unlimited context, for multi-step autonomous tasks a single-call agent cannot handle reliably.
• pydantic-ai-shields — drop-in guardrails: cost tracking, PII filtering, prompt injection detection, and secret redaction, built on Pydantic AI's native capabilities API.
• pydantic-ai-todo — hierarchical task planning and tracking for any Pydantic AI agent, with PostgreSQL multi-tenancy and an event system for webhooks and callbacks.
• pydantic-ai-backend — file storage and Docker-sandboxed execution backends for safe multi-user handling and isolated code execution within agent workflows.
• awesome-pydantic-ai — our curated practitioner index of the Pydantic AI ecosystem, maintained for teams actively building in this space.

None of these were built as portfolio projects. Each one was built because we needed it to ship a client system. We open-sourced each one because the problems they solve are not unique to our clients.

Our CEO Antoni Kozelski published a Pydantic guest article introducing pydantic-deep, and Vstorm client implementations appear directly on Pydantic's case studies.

"Vstorm haven't just adopted Pydantic AI, they've helped shape it – contributing extensions, pushing on the rough edges, and showing us where real-world agent systems break. That kind of collaboration is gold for an open source project." — Samuel Colvin , Founder & CEO, Pydantic

The triumphs: recognition from the Pydantic team #

That work is now formally endorsed. The Pydantic AI team's official capability library ( link) references Vstorm tooling across its capability matrix and states it plainly:

"Packages by vstorm-co are endorsed by the Pydantic AI team. We're working with them to upstream some of their implementations into this repo."

When a framework's maintainers move to absorb your implementations into their own codebase, that is the strongest form of recognition engineering offers.

"Whatever the vibe coding hype may suggest, building GenAI applications is still just engineering. […] We built Pydantic AI because no existing agent framework met our bar for developer ergonomics, engineering quality, and production readiness." — Samuel Colvin, September 2025

That shared orientation, production engineering over adoption velocity, is what made a deeper relationship between our two organisations a natural next step.

The allies: cooperation before the contract #

Before any formal agreement was in place, our working relationship had already taken shape. Vstorm engineers were contributing to the codebase. Client implementations were appearing on Pydantic's own case studies page. The guest article on pydantic.dev carried Vstorm's name as a practitioner voice in the ecosystem.

This reflects something important about how durable partnerships form in engineering-led organisations. Our alignment was demonstrated through real, practical work before it was formalised through paperwork. By the time the formal structure was proposed, both sides already understood what the other brought to the table.

"Vstorm are one of the strongest agent engineering teams we've collaborated with. Their work on and with Pydantic AI speaks for itself." — Samuel Colvin , Founder & CEO, Pydantic

The agreement: the formal partnership #

We are now proud to announce our formalised partnership with Pydantic as Vstorm becomes a primary agentic AI transformation partner and ambassador for organisations in the EU and UK building with Pydantic.

In practice, this means we are a trusted partner in delivering whichever parts of the Pydantic stack a client needs: type-safe agents with Pydantic AI, observability and evals with Pydantic Logfire, model routing and cost control via Pydantic AI Gateway (now part of Logfire), evaluation with Pydantic Evals, and any combination thereof.

The implementation dimension of the partnership is grounded in our extensive delivery record predating the formal agreement. And organisations evaluating Pydantic for production deployment can draw directly on that track record: such as the order completion agent built for Mixam, handling complex multi-step product logic at scale; the bilingual RAG chatbot built for the ARIJ Network, serving journalists across 22 countries; and the Hybrid Agent-Graph Architecture case study, which documents the architectural decisions that move a single-agent prototype to a production-grade multi-agent system.

These are the deployments the partnership is built on.

Vstorm as Pydantic ambassador #

We formalise Vstorm's commitment to building the Pydantic ecosystem in accepting the role of ambassador. This covers key three areas and activities to promote the framework across the industry:

Conferences

We will co-host and co-organise on-site conferences across Europe, beginning with events in the UK and Germany. These are not developer meetups, but designed for decision-makers: CTOs, VPs of Engineering, and heads of AI transformation at mid-market and upper mid-market organisations. The format will combine Pydantic's technical depth with Vstorm's applied production experience, giving attendees a grounded view of what the framework actually delivers in enterprise environments, drawn from real implementations rather than demonstrations.

Hackathons

We will run hackathons positioned as technical-depth events, where participants build on Pydantic AI frameworks from the ground up. These events will be run in cooperation with leading universities in the region, bringing academic rigour into contact with production engineering practice. The aim is to develop the next generation of engineers who build agentic systems the right way from the start, with type safety, observability, and production readiness as defaults rather than afterthoughts.

Training materials

We will co-create educational content with the Pydantic team, covering courses, webinars, and video content grounded in applied real-world experience. The material will draw directly from the client implementations, open-source tools, and engineering decisions that have shaped our production work over the past years. Where the Pydantic team brings framework depth, we bring the operational context of what practitioners actually encounter when they deploy these systems inside real organisations.

What this means for organisations in the EU and UK #

MIT's NANDA initiative published The GenAI Divide: State of AI in Business 2025 in August 2025, drawing on 150 leadership interviews, 350 employee surveys, and 300 public AI deployment analyses. Their finding: 95% of enterprise generative AI pilots fail to deliver measurable business impact. The identified root cause was not model quality but flawed enterprise integration.

The partnership between Vstorm and Pydantic addresses that failure mode at its source, because failure is rarely due to the model itself, but to all parts around it.

This is precisely what the Pydantic stack is built to cover: Pydantic AI for type-safe agents across 30+ model providers, Pydantic Logfire for AI observability and evals, Pydantic AI Gateway for model routing and cost control, and Pydantic Evals for measuring quality. It is a reliable end-to-end AI engineering stack built on open standards like OpenTelemetry, with no vendor lock-in.

Vstorm provides the production-grade agentic AI implementation depth, the open-source tooling, and the engineering track record to deploy that stack inside real infrastructure and transfer ownership to the client's team when the engagement ends.

For mid-market organisations in the EU and UK, the practical implication is access to a production-ready, end-to-end AI engineering stack, delivered by the agentic AI implementation partner with the deepest operational knowledge of it in the region. The goal, in every engagement, is a production system the client's team understands, can maintain, and can extend, running on open-source tooling which no single vendor controls.

Our TriStorm methodology governs the full journey from roadmap to deployed system. Our open-source contributions are available to any team building in this space.